Joe Rose, president of JBS Dev, pushes back against a widespread misconception blocking AI adoption. Companies often delay generative and agentic AI projects waiting for perfect datasets. This assumption wastes time and money.
Rose argues that imperfect data doesn't disqualify AI workloads. Real-world datasets contain gaps, inconsistencies, and noise. Waiting for perfection means projects stall indefinitely. Instead, organizations should deploy AI systems with messy data, then refine both the model and data quality iteratively.
This approach addresses what Rose calls "the AI last mile" – the gap between impressive model capabilities in labs and sustainable, cost-effective deployment in production. The last mile involves translating raw model performance into business value while managing real operational constraints.
The practical reality differs sharply from vendor narratives. High-quality training data improves results, but diminishing returns kick in quickly. A model trained on 70% clean data often outperforms competitors stuck preparing perfect datasets. Rose emphasizes that data cleanup happens faster once systems run on real workloads and surface actual problems.
Cost sustainability emerges as the critical factor most vendors gloss over. A model that costs $10 per inference query fails regardless of accuracy if the business case requires $1 per query. JBS Dev focuses on optimizing the economics alongside capability. This means choosing appropriate model sizes, implementing caching, batching requests, and fine-tuning for specific use cases rather than general performance.
Organizations attempting AI projects typically underestimate the work required after model selection. Integration, data pipeline development, monitoring, and continuous refinement consume 60-70% of project effort. Rose's position shifts focus from training perfection to operational reality.
The message matters for enterprises evaluating generative AI spending. Start with available data. Deploy incrementally. Measure real costs in production. Improve both the model and data